arch/x86: introduce postcar stage/phase

Certain chipsets don't have a memory-mapped boot media
so their code execution for stages prior to DRAM initialization
is backed by SRAM or cache-as-ram. The postcar stage/phase
handles the cache-as-ram situation where in order to tear down
cache-as-ram one needs to be executing out of a backing
store that isn't transient. By current definition, cache-as-ram
is volatile and tearing it down leads to its contents disappearing.
Therefore provide a shim layer, postcar, that's loaded into
memory and executed which does 2 things:

1. Tears down cache-as-ram with a chipset helper function.
2. Loads and runs ramstage.

Because those 2 things are executed out of ram there's no issue
of the code's backing store while executing the code that
tears down cache-as-ram. The current implementation makes no
assumption regarding cacheability of the DRAM itself. If the
chipset code wishes to cache DRAM for loading of the postcar
stage/phase then it's also up to the chipset to handle any
coherency issues pertaining to cache-as-ram destruction.

Change-Id: Ia58efdadd0b48f20cfe7de2f49ab462306c3a19b
Signed-off-by: Aaron Durbin <adurbin@chromium.org>
Reviewed-on: https://review.coreboot.org/14140
Tested-by: build bot (Jenkins)
Reviewed-by: Patrick Georgi <pgeorgi@google.com>
Reviewed-by: Furquan Shaikh <furquan@google.com>

Makefile.inc

@@ -84,7 +84,7 @@ subdirs-y += site-local
 
 #######################################################################
 # Add source classes and their build options
-classes-y := ramstage romstage bootblock smm smmstub cpu_microcode libverstage verstage
+classes-y := ramstage romstage bootblock postcar smm smmstub cpu_microcode libverstage verstage
 
 # Add dynamic classes for rmodules
 $(foreach supported_arch,$(ARCH_SUPPORTED), \

src/arch/x86/Kconfig

@@ -167,3 +167,8 @@ config ROMSTAGE_ADDR
 config VERSTAGE_ADDR
 	hex
 	default 0x2000000
+
+# Use the post CAR infrastructure for tearing down cache-as-ram
+# from a program loaded in ram and subsequently loading ramstage.
+config POSTCAR_STAGE
+	def_bool n

src/arch/x86/Makefile.inc

@@ -331,6 +331,7 @@ endif # CONFIG_ARCH_RAMSTAGE_X86_32 / CONFIG_ARCH_RAMSTAGE_X86_64
 
 ifeq ($(CONFIG_ARCH_ROMSTAGE_X86_32)$(CONFIG_ARCH_ROMSTAGE_X86_64),y)
 
+romstage-$(CONFIG_POSTCAR_STAGE) += postcar_loader.c
 romstage-y += cbmem.c
 romstage-y += boot.c
 
@@ -390,3 +391,27 @@ rmodules_x86_64-y += memmove.c
 endif
 
 endif # CONFIG_ARCH_RAMSTAGE_X86_32 / CONFIG_ARCH_RAMSTAGE_X86_64
+
+$(eval $(call create_class_compiler,postcar,x86_32))
+postcar-generic-ccopts += -D__POSTCAR__
+
+postcar-y += boot.c
+postcar-y += cbfs_and_run.c
+postcar-y += exit_car.S
+postcar-y += memset.c
+postcar-y += memcpy.c
+postcar-y += memmove.c
+postcar-y += memlayout.ld
+postcar-$(CONFIG_X86_TOP4G_BOOTMEDIA_MAP) += mmap_boot.c
+
+$(objcbfs)/postcar.debug: $$(postcar-objs)
+	@printf "    LINK       $(subst $(obj)/,,$(@))\n"
+	$(LD_postcar) $(LDFLAGS_postcar) -o $@ -L$(obj) $(COMPILER_RT_FLAGS_postcar) --whole-archive --start-group $(filter-out %.ld,$^) --no-whole-archive $(COMPILER_RT_postcar) --end-group -T $(call src-to-obj,postcar,src/arch/x86/memlayout.ld)
+
+$(objcbfs)/postcar.elf: $(objcbfs)/postcar.debug.rmod
+	cp $< $@
+
+cbfs-files-$(CONFIG_POSTCAR_STAGE)  += $(CONFIG_CBFS_PREFIX)/postcar
+$(CONFIG_CBFS_PREFIX)/postcar-file := $(objcbfs)/postcar.elf
+$(CONFIG_CBFS_PREFIX)/postcar-type := stage
+$(CONFIG_CBFS_PREFIX)/postcar-compression := none

src/arch/x86/exit_car.S

@@ -0,0 +1,106 @@
+/*
+ * This file is part of the coreboot project.
+ *
+ * Copyright 2016 Google Inc.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; version 2 of
+ * the License.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ */
+
+#include <cpu/x86/mtrr.h>
+#include <cpu/x86/cr.h>
+
+.section ".module_parameters", "aw", @progbits
+/* stack_top indicates the stack to pull MTRR information from. */
+stack_top:
+.long	0
+.long	0
+
+.text
+.global _start
+_start:
+	/* chipset_teardown_car() is expected to disable cache-as-ram. */
+	call	chipset_teardown_car
+
+	/* Enable caching if not already enabled. */
+	mov	%cr0, %eax
+	and	$(~(CR0_CD | CR0_NW)), %eax
+	mov	%eax, %cr0
+
+	/* Ensure cache is clean. */
+	invd
+
+	/* Set up new stack. */
+	mov	stack_top, %esp
+
+	/*
+	 * Honor variable MTRR information pushed on the stack with the
+	 * following layout:
+	 *
+	 * Offset: Value
+	 *   ...
+	 *   0x14: MTRR mask 0 63:32
+	 *   0x10: MTRR mask 0 31:0
+	 *   0x0c: MTRR base 0 63:32
+	 *   0x08: MTRR base 0 31:0
+	 *   0x04: Number of variable MTRRs to set
+	 *   0x00: Number of variable MTRRs to clear
+	 */
+
+	/* Clear variable MTRRs. */
+	pop	%ebx	/* Number to clear */
+	test	%ebx, %ebx
+	jz	2f
+	xor	%eax, %eax
+	xor	%edx, %edx
+	mov	$(MTRR_PHYS_BASE(0)), %ecx
+1:
+	wrmsr
+	inc	%ecx
+	wrmsr
+	inc	%ecx
+	dec	%ebx
+	jnz	1b
+2:
+
+	/* Set Variable MTRRs based on stack contents. */
+	pop	%ebx	/* Number to set. */
+	test	%ebx, %ebx
+	jz	2f
+	mov	$(MTRR_PHYS_BASE(0)), %ecx
+1:
+	/* Write MTRR base. */
+	pop	%eax
+	pop	%edx
+	wrmsr
+	inc	%ecx
+	/* Write MTRR mask. */
+	pop	%eax
+	pop	%edx
+	wrmsr
+	inc	%ecx
+
+	dec	%ebx
+	jnz	1b
+2:
+
+	/* Enable MTRR. */
+	mov	$(MTRR_DEF_TYPE_MSR), %ecx
+	rdmsr
+	/* Make default type uncacheable. */
+	and	$(~(MTRR_DEF_TYPE_MASK)), %eax
+	or	$(MTRR_DEF_TYPE_EN), %eax
+	wrmsr
+
+	/* Load and run ramstage. */
+	call	copy_and_run
+	/* Should never return. */
+1:
+	jmp	1b

src/arch/x86/include/arch/cpu.h

@@ -15,6 +15,7 @@
 #define ARCH_CPU_H
 
 #include <stdint.h>
+#include <stddef.h>
 #include <rules.h>
 
 /*
@@ -246,6 +247,42 @@ static inline void get_fms(struct cpuinfo_x86 *c, uint32_t tfms)
  * cache-as-ram.
  */
 void asmlinkage car_stage_entry(void);
+
+/*
+ * Support setting up a stack frame consisting of MTRR information
+ * for use in bootstrapping the caching attributes after cache-as-ram
+ * is torn down.
+ */
+
+struct postcar_frame {
+	uintptr_t stack;
+	uint32_t upper_mask;
+	int max_var_mttrs;
+	int num_var_mttrs;
+};
+
+/*
+ * Initialize postcar_frame object allocating stack size in cbmem
+ * with the provided size. Returns 0 on success, < 0 on error.
+ */
+int postcar_frame_init(struct postcar_frame *pcf, size_t stack_size);
+
+/*
+ * Add variable MTRR covering the provided range with MTRR type.
+ */
+void postcar_frame_add_mtrr(struct postcar_frame *pcf,
+				uintptr_t addr, size_t size, int type);
+
+/*
+ * Load and run a program that takes control of execution that
+ * tears down CAR and loads ramstage. The postcar_frame object
+ * indicates how to set up the frame. If caching is enabled at
+ * the time of the call it is up to the platform code to handle
+ * coherency with dirty lines in the cache using some mechansim
+ * such as platform_prog_run() because run_postcar_phase()
+ * utilizes prog_run() internally.
+ */
+void run_postcar_phase(struct postcar_frame *pcf);
 #endif
 
 #endif /* ARCH_CPU_H */

src/arch/x86/memlayout.ld

@@ -51,6 +51,8 @@ SECTIONS
 	/* Pull in the cache-as-ram rules. */
 	#include "car.ld"
 
+#elif ENV_POSTCAR
+	POSTCAR(32M, 1M)
 #endif
 }
 

src/arch/x86/postcar_loader.c

@@ -0,0 +1,115 @@
+/*
+ * This file is part of the coreboot project.
+ *
+ * Copyright 2016 Google Inc.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; version 2 of the License.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ */
+
+#include <arch/cpu.h>
+#include <cbmem.h>
+#include <console/console.h>
+#include <cpu/x86/msr.h>
+#include <cpu/x86/mtrr.h>
+#include <program_loading.h>
+#include <rmodule.h>
+
+static inline void stack_push(struct postcar_frame *pcf, uint32_t val)
+{
+	uint32_t *ptr;
+
+	pcf->stack -= sizeof(val);
+	ptr = (void *)pcf->stack;
+	*ptr = val;
+}
+
+int postcar_frame_init(struct postcar_frame *pcf, size_t stack_size)
+{
+	void *stack;
+	msr_t msr;
+
+	msr = rdmsr(MTRR_CAP_MSR);
+
+	stack = cbmem_add(CBMEM_ID_ROMSTAGE_RAM_STACK, stack_size);
+	if (stack == NULL) {
+		printk(BIOS_ERR, "Couldn't add %zd byte stack in cbmem.\n",
+			stack_size);
+		return -1;
+	}
+
+	pcf->stack = (uintptr_t)stack;
+	pcf->stack += stack_size;
+
+	pcf->upper_mask = (1 << (cpu_phys_address_size() - 32)) - 1;
+
+	pcf->max_var_mttrs = msr.lo & MTRR_CAP_VCNT;
+
+	pcf->num_var_mttrs = 0;
+
+	return 0;
+}
+
+void postcar_frame_add_mtrr(struct postcar_frame *pcf,
+				uintptr_t addr, size_t size, int type)
+{
+	size_t align;
+
+	if (pcf->num_var_mttrs >= pcf->max_var_mttrs) {
+		printk(BIOS_ERR, "No more variable MTRRs: %d\n",
+			pcf->max_var_mttrs);
+		return;
+	}
+
+	/* Determine address alignment by lowest bit set in address. */
+	align = addr & (addr ^ (addr - 1));
+
+	if (align < size) {
+		printk(BIOS_ERR, "Address (%lx) alignment (%zx) < size (%zx)\n",
+			addr, align, size);
+		size = align;
+	}
+
+	/* Push MTRR mask then base -- upper 32-bits then lower 32-bits. */
+	stack_push(pcf, pcf->upper_mask);
+	stack_push(pcf, ~(size - 1) | MTRR_PHYS_MASK_VALID);
+	stack_push(pcf, 0);
+	stack_push(pcf, addr | type);
+	pcf->num_var_mttrs++;
+}
+
+void run_postcar_phase(struct postcar_frame *pcf)
+{
+	struct prog prog =
+		PROG_INIT(PROG_UNKNOWN, CONFIG_CBFS_PREFIX "/postcar");
+	struct rmod_stage_load rsl = {
+		.cbmem_id = CBMEM_ID_AFTER_CAR,
+		.prog = &prog,
+	};
+
+	/*
+	 * Place the number of used variable MTRRs on stack then max number
+	 * of variable MTRRs supported in the system.
+	 */
+	stack_push(pcf, pcf->num_var_mttrs);
+	stack_push(pcf, pcf->max_var_mttrs);
+
+	if (prog_locate(&prog))
+		die("Failed to locate after CAR program.\n");
+	if (rmodule_stage_load(&rsl))
+		die("Failed to load after CAR program.\n");
+
+	/* Set the stack pointer within parameters of the program loaded. */
+	if (rsl.params == NULL)
+		die("No parameters found in after CAR program.\n");
+
+	*(uintptr_t *)rsl.params = pcf->stack;
+
+	prog_run(&prog);
+}

src/commonlib/Makefile.inc

@@ -2,12 +2,14 @@ bootblock-y += mem_pool.c
 verstage-y += mem_pool.c
 romstage-y += mem_pool.c
 ramstage-y += mem_pool.c
+postcar-y += mem_pool.c
 
 bootblock-y += region.c
 verstage-y += region.c
 romstage-y += region.c
 ramstage-y += region.c
 smm-y += region.c
+postcar-y += region.c
 
 ramstage-$(CONFIG_PLATFORM_USES_FSP1_1) += fsp1_1_relocate.c
 
@@ -16,8 +18,10 @@ verstage-y += cbfs.c
 romstage-y += cbfs.c
 ramstage-y += cbfs.c
 smm-y += cbfs.c
+postcar-y += cbfs.c
 
 bootblock-y += lz4_wrapper.c
 verstage-y += lz4_wrapper.c
 romstage-y += lz4_wrapper.c
 ramstage-y += lz4_wrapper.c
+postcar-y += lz4_wrapper.c

src/commonlib/include/commonlib/cbmem_id.h

@@ -20,6 +20,7 @@
 #define CBMEM_ID_ACPI		0x41435049
 #define CBMEM_ID_ACPI_GNVS	0x474e5653
 #define CBMEM_ID_ACPI_GNVS_PTR	0x474e5650
+#define CBMEM_ID_AFTER_CAR	0xc4787a93
 #define CBMEM_ID_AGESA_RUNTIME	0x41474553
 #define CBMEM_ID_AMDMCT_MEMINFO 0x494D454E
 #define CBMEM_ID_CAR_GLOBALS	0xcac4e6a3
@@ -71,6 +72,7 @@
 	{ CBMEM_ID_ACPI_GNVS,		"ACPI GNVS  " }, \
 	{ CBMEM_ID_ACPI_GNVS_PTR,	"GNVS PTR   " }, \
 	{ CBMEM_ID_AGESA_RUNTIME,	"AGESA RSVD " }, \
+	{ CBMEM_ID_AFTER_CAR,		"AFTER CAR"   }, \
 	{ CBMEM_ID_AMDMCT_MEMINFO,	"AMDMEM INFO" }, \
 	{ CBMEM_ID_CAR_GLOBALS,		"CAR GLOBALS" }, \
 	{ CBMEM_ID_CBTABLE,		"COREBOOT   " }, \

src/console/Makefile.inc

@@ -18,6 +18,11 @@ romstage-y += init.c console.c
 romstage-y += post.c
 romstage-y += die.c
 
+postcar-y += vtxprintf.c printk.c
+postcar-y += init.c console.c
+postcar-y += post.c
+postcar-y += die.c
+
 bootblock-$(CONFIG_BOOTBLOCK_CONSOLE) += vtxprintf.c printk.c
 bootblock-$(CONFIG_BOOTBLOCK_CONSOLE) += init.c console.c
 bootblock-y += die.c

src/cpu/x86/lapic/Makefile.inc

@@ -7,3 +7,4 @@ bootblock-y += boot_cpu.c
 verstage-y += boot_cpu.c
 romstage-y += boot_cpu.c
 ramstage-y += boot_cpu.c
+postcar-y += boot_cpu.c

src/include/console/console.h

@@ -47,7 +47,7 @@ void __attribute__ ((noreturn)) die(const char *msg);
 
 #define __CONSOLE_ENABLE__ \
 	((ENV_BOOTBLOCK && CONFIG_BOOTBLOCK_CONSOLE) || \
-	ENV_VERSTAGE || ENV_ROMSTAGE || ENV_RAMSTAGE || \
+	ENV_VERSTAGE || ENV_ROMSTAGE || ENV_RAMSTAGE || ENV_POSTCAR || \
 	(ENV_SMM && CONFIG_DEBUG_SMI))
 
 #if __CONSOLE_ENABLE__

src/include/memlayout.h

@@ -160,6 +160,18 @@
 	#define OVERLAP_VERSTAGE_ROMSTAGE(addr, size) ROMSTAGE(addr, size)
 #endif
 
+#if ENV_POSTCAR
+	#define POSTCAR(addr, sz) \
+		SYMBOL(postcar, addr) \
+		_epostcar = _postcar + sz; \
+		_ = ASSERT(_eprogram - _program <= sz, \
+			STR(Aftercar exceeded its allotted size! (sz))); \
+		INCLUDE "postcar/lib/program.ld"
+#else
+	#define POSTCAR(addr, sz) \
+		REGION(postcar, addr, sz, 1)
+#endif
+
 #define WATCHDOG_TOMBSTONE(addr, size) \
 	REGION(watchdog_tombstone, addr, size, 4) \
 	_ = ASSERT(size == 4, "watchdog tombstones should be exactly 4 byte!");

src/include/rules.h

@@ -26,6 +26,7 @@
 #define ENV_SMM 0
 #define ENV_VERSTAGE 0
 #define ENV_RMODULE 0
+#define ENV_POSTCAR 0
 #define ENV_STRING "bootblock"
 
 #elif defined(__ROMSTAGE__)
@@ -35,6 +36,7 @@
 #define ENV_SMM 0
 #define ENV_VERSTAGE 0
 #define ENV_RMODULE 0
+#define ENV_POSTCAR 0
 #define ENV_STRING "romstage"
 
 #elif defined(__SMM__)
@@ -44,6 +46,7 @@
 #define ENV_SMM 1
 #define ENV_VERSTAGE 0
 #define ENV_RMODULE 0
+#define ENV_POSTCAR 0
 #define ENV_STRING "smm"
 
 #elif defined(__VERSTAGE__)
@@ -53,6 +56,7 @@
 #define ENV_SMM 0
 #define ENV_VERSTAGE 1
 #define ENV_RMODULE 0
+#define ENV_POSTCAR 0
 #define ENV_STRING "verstage"
 
 #elif defined(__RAMSTAGE__)
@@ -62,6 +66,7 @@
 #define ENV_SMM 0
 #define ENV_VERSTAGE 0
 #define ENV_RMODULE 0
+#define ENV_POSTCAR 0
 #define ENV_STRING "ramstage"
 
 #elif defined(__RMODULE__)
@@ -71,8 +76,19 @@
 #define ENV_SMM 0
 #define ENV_VERSTAGE 0
 #define ENV_RMODULE 1
+#define ENV_POSTCAR 0
 #define ENV_STRING "rmodule"
 
+#elif defined(__POSTCAR__)
+#define ENV_BOOTBLOCK 0
+#define ENV_ROMSTAGE 0
+#define ENV_RAMSTAGE 0
+#define ENV_SMM 0
+#define ENV_VERSTAGE 0
+#define ENV_RMODULE 0
+#define ENV_POSTCAR 1
+#define ENV_STRING "postcar"
+
 #else
 /*
  * Default case of nothing set for random blob generation using
@@ -86,18 +102,21 @@
 #define ENV_SMM 0
 #define ENV_VERSTAGE 0
 #define ENV_RMODULE 0
+#define ENV_POSTCAR 0
 #define ENV_STRING "UNKNOWN"
 #endif
 
-/* For romstage and ramstage always build with simple device model, ie.
+/* For pre-DRAM stages and post-CAR always build with simple device model, ie.
- * PCI, PNP and CPU functions operate without use of devicetree.
+ * PCI, PNP and CPU functions operate without use of devicetree. The reason
+ * post-CAR utilizes __SIMPLE_DEVICE__ is for simplicity. Currently there's
+ * no known requirement that devicetree would be needed during that stage.
  *
  * For ramstage individual source file may define __SIMPLE_DEVICE__
  * before including any header files to force that particular source
  * be built with simple device model.
  */
 
-#if defined(__PRE_RAM__) || defined(__SMM__)
+#if defined(__PRE_RAM__) || ENV_SMM || ENV_POSTCAR
 #define __SIMPLE_DEVICE__
 #endif
 

src/lib/Makefile.inc

@@ -131,13 +131,16 @@ ramstage-$(CONFIG_ACPI_NHLT) += nhlt.c
 
 romstage-y += cbmem_common.c
 romstage-y += imd_cbmem.c
+romstage-y += imd.c
 
 ramstage-y += cbmem_common.c
 ramstage-y += imd_cbmem.c
-
-romstage-y += imd.c
 ramstage-y += imd.c
 
+postcar-y += cbmem_common.c
+postcar-y += imd_cbmem.c
+postcar-y += imd.c
+
 bootblock-y += hexdump.c
 ramstage-y += hexdump.c
 romstage-y += hexdump.c
@@ -167,12 +170,14 @@ romstage-y += version.c
 ramstage-y += version.c
 smm-y += version.c
 verstage-y += version.c
+postcar-y += version.c
 
 $(call src-to-obj,bootblock,$(dir)/version.c) : $(obj)/build.h
 $(call src-to-obj,romstage,$(dir)/version.c) : $(obj)/build.h
 $(call src-to-obj,ramstage,$(dir)/version.c) : $(obj)/build.h
 $(call src-to-obj,smm,$(dir)/version.c) : $(obj)/build.h
 $(call src-to-obj,verstage,$(dir)/version.c) : $(obj)/build.h
+$(call src-to-obj,postcar,$(dir)/version.c) : $(obj)/build.h
 
 romstage-y += bootmode.c
 ramstage-y += bootmode.c
@@ -183,9 +188,27 @@ romstage-y += halt.c
 ramstage-y += halt.c
 smm-y += halt.c
 
+postcar-y += bootmode.c
+postcar-y += boot_device.c
+postcar-y += cbfs.c
+postcar-$(CONFIG_COMMON_CBFS_SPI_WRAPPER) += cbfs_spi.c
+postcar-y += delay.c
+postcar-y += fmap.c
+postcar-y += gcc.c
+postcar-y += halt.c
+postcar-y += libgcc.c
+postcar-$(CONFIG_COMPRESS_RAMSTAGE) += lzma.c lzmadecode.c
+postcar-y += memchr.c
+postcar-y += memcmp.c
+postcar-y += prog_loaders.c
+postcar-y += prog_ops.c
+postcar-y += rmodule.c
+postcar-$(CONFIG_COLLECT_TIMESTAMPS) += timestamp.c
+
 # Use program.ld for all the platforms which use C fo the bootblock.
 bootblock-$(CONFIG_C_ENVIRONMENT_BOOTBLOCK) += program.ld
 
+postcar-y += program.ld
 romstage-y += program.ld
 ramstage-y += program.ld
 verstage-y += program.ld

src/lib/program.ld

@@ -80,7 +80,14 @@
 	. = ALIGN(ARCH_CACHELINE_ALIGN_SIZE);
 	_data = .;
 
-#if ENV_RMODULE
+/*
+ * The postcar phase uses a stack value that is located in the relocatable
+ * module section. While the postcar stage could be linked like smm and
+ * other rmodules the postcar stage needs similar semantics of the more
+ * traditional stages in the coreboot infrastructure. Therefore it's easier
+ * to specialize this case.
+ */
+#if ENV_RMODULE || ENV_POSTCAR
 	_rmodule_params = .;
 	KEEP(*(.module_parameters));
 	_ermodule_params = .;